Biomarkers of mammographic density in premenopausal women

The MTC has been described in detail elsewhere [22]. In brief, this cohort study started with enrollment of 27,979 female teachers from the Mexican states of Jalisco and Veracruz in 2006, before the recruitment was extended to other states of Mexico in 2008 to reach a total of 115,314 female participants. With the aim to characterize risk factors related to cancer and other chronic diseases, women were administered lifestyle and dietary questionnaires. In 2007, a subsample of 2,045 women from the regions of Jalisco and Veracruz participated in a clinical examination that included an interview, anthropometric measurements by trained personnel, a mammogram, and biological samples collection. All participants gave informed consent for future use of biological specimens and questionnaire data. The International Agency for Research on Cancer (IARC) Ethics Committee as well as the Research Ethics Committee at the National Institute of Public Health in Cuernavaca, Mexico, approved the current project.

Trained nurses collected fasting blood samples (25 mL). Plasma, serum, erythrocytes, and buffy coat were separated by centrifugation at 2500 rpm for 10 min in a refrigerated centrifuge (4 °C) and aliquoted into several cryotubes at field work site, within 30 min after blood collection. Samples were frozen and kept in liquid nitrogen at the National Institute of Public Health in Cuernavaca, Mexico, until shipment to IARC, where they were stored at − 80 °C until metabolomics analyses were run.

Out of the 2045 women who underwent clinical examination, 230 with missing information on metabolic syndrome components were excluded (based on inclusion criteria for a previous study [23]), 67 because of unknown menopausal status, and 624 who were postmenopausal at the time of their mammogram.

Menopausal status was defined as follows: premenopausal if women menstruated at least once over the 12 months prior to recruitment, postmenopausal if women had (1) no menstruation over the last 12 months prior to the clinical examination or (2) surgical menopause (reported bilateral oophorectomy or reported unknown surgery) and were over 48 years old (mean age at menopause in Mexico [24]).

Further selection was based on commonly used [25] breast density categories (< 10%, 10 to < 25%, 25 to < 50%, ≥ 50%): women from each group of breast density were randomly selected proportionally to the size of the group, among non-users of oral contraceptive at blood donation. A total of 35 women were selected for the first group, 158 for the second, 247 for the third, and 160 for the last group. Among the 600 women whose samples were analyzed for targeted metabolomics, 1 woman was excluded because she had no biological sample, 7 women were excluded because they were older than 55 but declared to be premenopausal, and 19 women for whom measured BMI was not available at the time of clinical evaluation. Our final population included 573 women.

A radiology technician performed mammography in both Jalisco and Veracruz. Cranio-caudal views were taken on each breast. Analog films were digitized using an Astra 2400S (Umax, Fremont, CA). Mammograms of both states were combined, and mammographic density was measured by a single observer on the cranio-caudal view of the left breast using Mamgr, a computer-assisted program developed at the Department of Epidemiology and Population Health, London School of Hygiene and Tropical Medicine, and based on previously reported mammographic density assessment methods [26]. In a validation study, intraclass correlation coefficient between mammographic density measurements performed with the Mamgr software versus with the Cumulus program was 0.87 (n = 100 mammograms), while the intra-observer intraclass correlation coefficient was 0.84 (n = 108 duplicates mammograms) [27]. Percent mammographic density was automatically calculated as the percent of dense pixels within the breast area. Non-dense area was calculated by subtracting the dense area from the total breast area. We converted absolute dense and non-dense area values to cm² according to pixel size used in digitalization.

Percent of samples out of the measurable range and coefficients of variation for metabolites included in the analysis (median = 6.0%, interquartile range = 2.1%) are shown in Supplementary Table 1.

Data on dietary habits were collected through a 139-item food frequency questionnaire (details previously published [28]). Information on frequency of consumption and portion size were used to calculate nutrients and energy intakes using the United States Department of Agriculture food-composition database and the Mexican National Health and Nutrition Survey database. Three dietary patterns were identified by principal component analysis (“Fruits and Vegetables,” “Western,” “Modern Mexican”) [28]. Intakes and frequency were also used to estimate the Healthy Eating Index (HEI) 2015 total score [29].

Insulin-like growth factor 1 (IGF-1), IGF binding protein 3 (IGFBP-3), C-peptide, C reactive protein (CRP), leptin, and adiponectin analyses were performed in the laboratory of the Biomarkers Group at IARC [30]. Serum IGF-I, IGFBP3, and C-peptide concentrations were measured by immunoradiometric assays by Beckmann Coulter (Marseille, France) [13]. Leptin was measured by a radioimmunoassay from Linco (Millipore, Billerica, MA, USA), while adiponectin and CRP were measured using an enzyme-linked immunoassay by R&D (R&D Systems, Europe, Lille, France) [15].

Triglycerides, total and HDL cholesterol, and glucose were measured on fasting plasma blood samples at the Endocrinology and Metabolism Laboratory at the National Institute of Nutrition and Medical Sciences using standard assays. Glucose was measured via the automatized glucose oxidase method; triglycerides and HDLs were measured using enzymatic hydrolysis in an automatic analyzer with a tungsten lamp (Prestige 24i, Tokyo Boeki Medical System LTD). Number of metabolic syndrome components was defined according to the harmonized definition [31] (waist circumference ≥ 88 cm, triglyceride levels ≥ 150 mg/dL, HDL cholesterol levels < 50 mg/dL, systolic blood pressure > 130 mmHg or diastolic blood pressure > 85 mmHg, and glucose levels ≥ 100 mg/dL) (details previously published [23]).

Descriptive analyses were performed for selected characteristics of the population using mean and standard deviation (continuous variables) or frequency (categorical variables). Partial Pearson’s correlation coefficients, adjusted for age (where appropriate), state, and batch were computed for metabolites, measures of mammographic density, age, and BMI. Percent mammographic density was the primary outcome of this analysis, while dense area and non-dense area were examined as secondary outcomes after log-transformation to better approximate normality and homoscedasticity of the residuals. To account for analytical batch, residuals of log-transformed and standardized metabolites concentrations were obtained from linear models with random effect for analytical batches. These residuals were used as dependent variables in multiple linear regression testing associations with the different outcomes.

All models were adjusted for a priori selected breast cancer risk factors that included: age (continuous), BMI (continuous), age at menarche (< 12, 12, 13, ≥ 14 years, missing), family history of cancer (yes, no), history of benign breast disease (yes, no), use of oral contraceptive (ever, never), number of full-term pregnancies (0, 1, 2, 3, ≥ 4, missing), age at first full-term pregnancy (nulliparous, < 20, 20–25, 25–30, ≥ 30, missing), breastfeeding (nulliparous, no breastfeeding, < 6 months, 6–12 months, 12–24 months, ≥ 24 months, missing), alcohol intake (0, 0.1 drinks/day, 0.1–0.2 drinks/day, ≥ 0.2 drinks/day, missing), smoking status (never, past, current, missing), socioeconomic status (low, medium, high, missing), and physical activity (continuous). A missing category was created for all variables, except for physical activity where the only missing value was imputed to the median. Multiple tests were addressed using permutation minP-adjustment of P values to account for the dependencies between tests [32].

For metabolites associated with percent mammographic density after correction for multiple testing, adjusted means of percent mammographic density were estimated by quartile of metabolite. For test of linear trend, participants were assigned the median value of exposure in each quartile and we modeled the corresponding variable as a continuous term. Analyses were further stratified by BMI (</> 27.4 kg/m² (median)) and interaction with the dichotomized variable was tested for each metabolite by including an interaction term in the model. Adjusted means were examined by quartiles of metabolite in each group, and BMI (continuous) was included as an adjustment variable in each model.

To examine the robustness of the observed associations, additional exploratory analyses were conducted using a bootstrapped least absolute shrinkage and selection operator (LASSO) regression approach [33, 34]: metabolic signatures of the percent mammographic density were obtained via simple cross-validated LASSO, which efficiently selects the most predictive variables in high-dimensional sets of potential predictors. This approach was then applied to 200 bootstrap samples to determine which metabolites were most frequently included in the signature.

To provide a better understanding of the metabolites associated with percent mammographic density, a variety of lifestyle, dietary, anthropometric, and metabolic factors already available in the study population were investigated in separate models in relation to plasma concentration of retained metabolites. Adjusted mean concentrations of metabolites of interest (residuals on analytical batch) were estimated across categories of each variable after excluding participants with missing values, adjusting for age and state. All variables previously listed as covariates in the main analyses were examined using similar categories or tertiles for variables initially included as continuous. In addition to these variables, we investigated waist circumference (tertiles), hip circumference (tertiles), high blood pressure (yes, no), circulating leptin, adiponectin, leptin/adiponectin ratio, IGF-1, IGFBP-3, C-peptide, CRP (tertiles of log-transformed concentration regressed on respective analytical batches), total cholesterol (tertiles), HDL cholesterol (tertiles), total cholesterol/HDL cholesterol ratio (tertiles), triglycerides (tertiles), glucose (tertiles), and number of criteria for determination of metabolic syndrome. The following nutritional factors were also examined (tertiles): total daily energy intake, protein, carbohydrate, starch, sugar, fibers, lipid, fatty acids (total, trans, saturated, monounsaturated, polyunsaturated) intakes (as residuals on total energy intake), glycemic index (GI) and glycemic load (GL), dietary patterns (“Fruits and Vegetables,” “Western,” “Modern Mexican”), and the HEI score. Heterogeneity of means across categories was assessed by F test from analyses of variance for all 46 variables, and P values were corrected for multiple tests with a Bonferroni correction (P < 0.001(0.05/46)). When significant heterogeneity was detected, linear trend across ordinal categories was further tested by assigning the median value of each category to participants and including the variable as a continuous term in a linear regression model.

All statistical tests were two-sided. Analyses were performed using SAS 9.4 (SAS Institute, Cary, NC) and R Studio (packages NPC [35] and glmnet [36]).

Associations of metabolites and percent mammographic density before correction for multiple tests are shown in Figure 1A. After correction of P values for multiple tests, only two inverse associations remained borderline statistically significant, SM C16:1 (minP P value = 0.05) and PC ae C30:2 (minP P value = 0.05) (Figure 1B). P values for tests of linear trends across quartiles of metabolites were < 0.01 for both SM C16:1 and PC ae C30:2 (Table 2). For SM C16:1, mean percent mammographic density was 38.1 (95% confidence interval (CI) 32.2-44.0) in first quartile and 32.0 (26.3–37.8) in last quartile. For PC ae C30:2, mean percent mammographic density was 40.7 (34.9–46.5) in the first quartile and 35.0 (29.2–40.8) in last quartile.

For non-dense area, after correction for multiple tests, 16 metabolites remained associated with non-dense area: 12 PCs (ae C30:2, ae C40:2, ae C40:5, ae C32:2, aa C36:0, aa C28:1, ae C42:2, ae C42:5, ae C38:3, ae C40:4, aa C40:3, aa C42:4) and 4 SMs (C18:1, C16:1, C20:2, OH-C14:1) (Supplementary Figure 2), with P values for tests of linear trends across quartiles all ≤0.02 (data not shown).

When assessing the associations of percent mammographic density with SM C16:1 and PC ae C30:2 across BMI strata (Table 2), no statistically significant interaction was detected for any of the metabolites (P interaction≥0.31). Of note, a linear inverse association with percent mammographic density was observed among quartiles of SM C16:1 (P trend< 0.01) in women with a BMI lower than 27.4 kg/m² (n = 286), but not in women with a BMI ≥ 27.4 kg/m² (n = 287, P trend = 0.07). For PC ae C30:2, no statistically significant linear trend across quartiles was observed among women with a BMI < 27.4 kg/m² (P trend = 0.09), while a significant inverse association was observed among women with a BMI ≥ 27.4 kg/m² (P trend = < 0.01). There was no evidence of a statistical interaction between any of the other metabolites and BMI (dichotomized by median) (all P interaction ≥0.07) for percent mammographic density, and none of the other metabolites was associated with percent mammographic density in any of the BMI strata after adjustment for multiple tests (data not shown).

In exploratory analyses using cross-validated bootstrapped LASSO, SM C16:1 was the most frequently (in 84% of bootstrap samples) identified metabolite. In contrast, PC ae C30:2 was selected in the signature in only 18% of bootstrap samples.

Of all potential correlates investigated, only those for which a statistically significant association was detected after correction of P values (from F-tests) for multiple tests are shown in Figures 2 and 3.

Figure 2 shows adjusted mean concentrations of SM C16:1 (residuals) across categories of variables for which a statistically significant association was detected after correction of P values (from F-tests) for multiple tests and shows P for linear trend (all < 0.01) for these variables: leptin, adiponectin, total cholesterol, HDL cholesterol (direct associations), total/HDL cholesterol ratio, triglycerides, glucose, and number of components for metabolic syndrome (inverse associations).

For PC ae C30:2 (Figure 3), significant heterogeneity after correction for multiple tests of P values from F-tests was observed across categories of C-peptide, triglycerides, number of components for metabolic syndrome (inverse associations), total cholesterol, HDL cholesterol, and daily protein intake (direct associations).

No statistical interaction was observed with state of residence for the two metabolites. Factors associated with SM C16:1 in the whole population remained associated at a statistically significant level in stratified analyses, except for leptin and total cholesterol/HDL cholesterol ratio, for which the associations were no longer significant after Bonferroni correction in both states, and glucose, which was no longer significant in Jalisco after Bonferroni correction. Total cholesterol, HDL cholesterol, and number of components of metabolic syndrome remained associated with PC ae C30:2 after Bonferroni correction in each state, while the association with C-peptide and triglycerides were significant in Jalisco only, and protein intake was not associated with PC ae C30:2 in any of the states.